From d664315ae896174b49b1feadea75a26dcfd209f2 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Thu, 24 Aug 2017 07:55:15 +0000
Subject: [PATCH] IPRA: Don't assume called function is first call operand

Fixes not finding the called global for AMDGPU
call pseudoinstructions, which prevented IPRA
from doing much.

llvm-svn: 311637
---
 llvm/lib/CodeGen/RegUsageInfoPropagate.cpp | 28 +++++---
 llvm/test/CodeGen/AMDGPU/ipra.ll           | 84 +++++++++++++++++++++-
 2 files changed, 103 insertions(+), 9 deletions(-)

diff --git a/llvm/lib/CodeGen/RegUsageInfoPropagate.cpp b/llvm/lib/CodeGen/RegUsageInfoPropagate.cpp
index a39c9457b3a4..f6d45067816a 100644
--- a/llvm/lib/CodeGen/RegUsageInfoPropagate.cpp
+++ b/llvm/lib/CodeGen/RegUsageInfoPropagate.cpp
@@ -88,6 +88,19 @@ void RegUsageInfoPropagationPass::getAnalysisUsage(AnalysisUsage &AU) const {
   MachineFunctionPass::getAnalysisUsage(AU);
 }
 
+// Assumes call instructions have a single reference to a function.
+static const Function *findCalledFunction(const Module &M, MachineInstr &MI) {
+  for (MachineOperand &MO : MI.operands()) {
+    if (MO.isGlobal())
+      return dyn_cast<Function>(MO.getGlobal());
+
+    if (MO.isSymbol())
+      return M.getFunction(MO.getSymbolName());
+  }
+
+  return nullptr;
+}
+
 bool RegUsageInfoPropagationPass::runOnMachineFunction(MachineFunction &MF) {
   const Module *M = MF.getFunction()->getParent();
   PhysicalRegisterUsageInfo *PRUI = &getAnalysis<PhysicalRegisterUsageInfo>();
@@ -118,15 +131,14 @@ bool RegUsageInfoPropagationPass::runOnMachineFunction(MachineFunction &MF) {
         Changed = true;
       };
 
-      MachineOperand &Operand = MI.getOperand(0);
-      if (Operand.isGlobal())
-        UpdateRegMask(cast<Function>(Operand.getGlobal()));
-      else if (Operand.isSymbol())
-        UpdateRegMask(M->getFunction(Operand.getSymbolName()));
+      if (const Function *F = findCalledFunction(*M, MI)) {
+        UpdateRegMask(F);
+      } else {
+        DEBUG(dbgs() << "Failed to find call target function\n");
+      }
 
-      DEBUG(dbgs()
-            << "Call Instruction After Register Usage Info Propagation : \n");
-      DEBUG(dbgs() << MI << "\n");
+      DEBUG(dbgs() << "Call Instruction After Register Usage Info Propagation : "
+            << MI << '\n');
     }
   }
 
diff --git a/llvm/test/CodeGen/AMDGPU/ipra.ll b/llvm/test/CodeGen/AMDGPU/ipra.ll
index 53752b24f406..803855cd0327 100644
--- a/llvm/test/CodeGen/AMDGPU/ipra.ll
+++ b/llvm/test/CodeGen/AMDGPU/ipra.ll
@@ -1,4 +1,4 @@
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -verify-machineinstrs -enable-ipra < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -verify-machineinstrs -enable-ipra -amdgpu-sroa=0 < %s | FileCheck -check-prefix=GCN %s
 
 ; Kernels are not called, so there is no call preserved mask.
 ; GCN-LABEL: {{^}}kernel:
@@ -9,4 +9,86 @@ entry:
   ret void
 }
 
+; GCN-LABEL: {{^}}func:
+; GCN: ; NumVgprs: 8
+define void @func() #1 {
+  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}"() #0
+  ret void
+}
+
+; GCN-LABEL: {{^}}kernel_call:
+; GCN-NOT: buffer_store
+; GCN-NOT: buffer_load
+; GCN-NOT: readlane
+; GCN-NOT: writelane
+; GCN: flat_load_dword v8
+; GCN: s_swappc_b64
+; GCN-NOT: buffer_store
+; GCN-NOT: buffer_load
+; GCN-NOT: readlane
+; GCN-NOT: writelane
+; GCN: flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, v8
+
+; GCN: ; NumSgprs: 37
+; GCN: ; NumVgprs: 9
+define amdgpu_kernel void @kernel_call() #0 {
+  %vgpr = load volatile i32, i32 addrspace(1)* undef
+  tail call void @func()
+  store volatile i32 %vgpr, i32 addrspace(1)* undef
+  ret void
+}
+
+; GCN-LABEL: {{^}}func_regular_call:
+; GCN-NOT: buffer_store
+; GCN-NOT: buffer_load
+; GCN-NOT: readlane
+; GCN-NOT: writelane
+; GCN: flat_load_dword v8
+; GCN: s_swappc_b64
+; GCN-NOT: buffer_store
+; GCN-NOT: buffer_load
+; GCN-NOT: readlane
+; GCN-NOT: writelane
+; GCN: flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, v8
+
+; GCN: ; NumSgprs: 32
+; GCN: ; NumVgprs: 9
+define void @func_regular_call() #1 {
+  %vgpr = load volatile i32, i32 addrspace(1)* undef
+  tail call void @func()
+  store volatile i32 %vgpr, i32 addrspace(1)* undef
+  ret void
+}
+
+; GCN-LABEL: {{^}}func_tail_call:
+; GCN: s_waitcnt
+; GCN-NEXT: s_getpc_b64 s[6:7]
+; GCN-NEXT: s_add_u32 s6,
+; GCN-NEXT: s_addc_u32 s7,
+; GCN-NEXT: s_setpc_b64 s[6:7]
+
+; GCN: ; NumSgprs: 32
+; GCN: ; NumVgprs: 8
+define void @func_tail_call() #1 {
+  tail call void @func()
+  ret void
+}
+
+; GCN-LABEL: {{^}}func_call_tail_call:
+; GCN: flat_load_dword v8
+; GCN: s_swappc_b64
+; GCN: flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, v8
+; GCN: s_setpc_b64
+
+; GCN: ; NumSgprs: 32
+; GCN: ; NumVgprs: 9
+define void @func_call_tail_call() #1 {
+  %vgpr = load volatile i32, i32 addrspace(1)* undef
+  tail call void @func()
+  store volatile i32 %vgpr, i32 addrspace(1)* undef
+  tail call void @func()
+  ret void
+}
+
 attributes #0 = { nounwind }
+attributes #1 = { nounwind noinline }